Literature Review: Machine Learning in Cancer Diagnoses

Literature review: Machine Learning in Cancer
Diagnoses
1.1 Machine Learning
Machine learning (ML) is a study of mathematical algorithms and models that
computer systems use to perform a particular task to improve their performance
gradually (1).
The future of clinical testing is on the threshold of a significant revolution due to the
collaboration of advanced innovations and technologies that offer new perspectives on
big data digital sources, enormous computing power that helps in identifying useful
patterns in clinical data by employing the use of artificial intelligence and machine
learning technologies and finally, authority bodies that are already issuing guidelines
and taking on the change brought by these collaborations (16). Clinical testing has
remained stagnant over the past 30 years. Some of the reasons attributed to the
stagnation include:
a) The lack of appropriate bio-medical sources of data and advanced data
analytics tools that can be used to analyze data and come up with a useful
hypothesis that could stir and motivate the generation of innovative treatments
and therapies (17).
b) Reservations on the part of the authorities and regulatory bodies on risk
aversion, requirements, and uncertainty regarding the rapidly changing yet
unproved innovations like wireless healthcare monitoring and machine
learning (23).
Thus, clinical testing of new bio-medical cures for usefulness and safety purposes will
also require new approaches since there is evidence that the current treatments only
work for a specified small number of persons (29). The implementation of
innovations such as the next-generation sequencing has greatly enhanced both the
possibility of developing customized treatments and upgraded our comprehension of
disease mechanisms in large groups of patients (31).
Another significant difficulty in clinical testing and development is associated with
results reporting of many everyday clinical trials of natural healthcare treatment
effects that may not directly be seen as making personalized treatment decisions at the
normal point-of-care. The good news is that there are promising methodologies to
combat the challenge (34). For instance, the use of modernized processes, refining

safety and efficiency while reducing the toxicity and confrontational trials, the
exploitation of new clinical endpoints and treatment biomarkers that can be easily
monitored like circulating tumor DNA and finally, more significant insights inpatient
monitoring by the use of low-cost imaging and sensors (36). Standardized, secured,
and improved methods of collecting EHR data may be used as a reliable source of
data and medical evidence that may be useful in organizing clinical tests at the point-
of-care and should function to advance the clinical testing and development process.
Machine learning concepts and computer vision have transformed numerous traits of
the perception of human vision in identifying clinical patterns and sequences that are
meaning, for instance, in computer imaging data (24). Neural networks have so far
been applied in various functions, including classification, prediction of clinical sets
of data, medical image segmentation, and generation (30). Primarily, bio-medical
technology firms, technology corporations, and academic research laboratories have
all been researching the use of machine learning and artificial intelligence in the
following three main fronts.
(a) The development of in-depth learning methodologies on multimodal sources
of data like the combination of clinical data and genomic data to be able to
come up with new predictive frameworks and models (39).
(b) The use of segmentation and pattern recognition technologies on medical
images like internal organs, retina scans, and body scans to enable speedy
diagnosis and monitoring of the progress of the disease. Also, the use of
generative innovations and algorithms for computational augmentation of
current clinical testing data and imaging data sets (47).
(c) The implementation of machine learning algorithms to predict the
pharmaceutical characteristics of targets and molecular compounds for drug
development and discovery (48).
There have been very few reported cases of successful use of machine learning in
spite of all of the above propositions to implement machine learning in the
acceleration of medical studies and clinical testing. These lack in implementation of
machine learning can be associated with a number of reasons:
a) A short time has passed since the discovery of relevant technologies.
b) There exists a deficiency in the existing machine learning frameworks and
models and computer science deep-learning in the generation of more credible
and reliable and big data sets in the medical field.

Other minor factors that hinder the development and adoption of machine learning
and artificial intelligence methodologies in clinical testing (44), including:
a) Privacy, legal and ethical considerations in data sharing
b) The scarcity of big data that is of high-quality
c) Nascent rules, regulations, and laws
Research indicates that substitute learning systems to deep learning (28) have been
proposed but lack a full reception (11). These alternatives depend on the human brain
and its neocortex and only learn from fewer models. In the recent past, there have
been publications in the areas of applications of Deep Neural Networks in imaging
data, the implementation of Natural Language Processing to HER, pharmaceutical
features of compounds, clinical testing, diagnosis and treatments, and computer vision
uses in medical imaging.
1.2 Classifier Development
Many classifiers have been used in the literature for cancer diagnosis, in this section,
most of the significant classifiers are covered including Artificial Neural Nets
(ANNs), Support Vector Machines (SVMs), Decision Trees, Random Forests, and
AdaBoost along with its variants.
1.3 Artificial Neural Networks
Artificial Neural Networks (ANNs) are considered more effective in the diagnosis of
cancer as compared to traditional techniques. This is because ANNs significantly
decrease the need for there to be an interpretation of results obtained from imaging
methodologies and invasive procedures (2).
In addition, ANNs have been trained and thus can analyze individual diagnosis and
treatment procedures with high accuracy that is equivalent to that of an experienced
medical practitioner (4). Hence, the use of ANNs helps both the patient and the
medical personnel in coming up with the best and informed decisions.

Figure 1: ANN structure (22)
Early before an individual is diagnosed with cancer, ANNs may be employed to offer
insight on their vulnerability to the disease (6). The authors came up with ANNs
framework model that was used to project the probability of a person to develop the
cancer of the breast (8). The model analyzed nutrients, genes, interaction, and
demographic indicators. Its model had a 94.2 percent accuracy and was used to come
up with variables that contributed the most to breast cancer (10). ANNs have a
diagnostic utility function and thus makes it an efficient artificial intelligence
technique to handle big data (12).
In the same way, a different study suggested the use of genetic techniques to model
for evolving ANNs such that the neural networks can adapt in the same manner as
natural evolutions happened (15). Eventually, the model would be employed in
computer-aided mammographic mass detection.
After ANNs use in cancer diagnosis, they are also used to improve the cancer
treatment schemes for patients (27). Their contribution to improving these schemes is
unmatched to any physician. A study indicates that 15 years ago, artificial intelligence
technologies were used to show a projection of relapse in cancer of the bladder (50)
with an estimated accuracy rate of between 88 and 95 percent. Also, a systematic
literature review issued in 2006 shows that 21 out of 27 cancer clinical tests that used
ANNs showed benefits in using the ANNs technique (35).